RepeatZero — clearing the repeating bulk from queue

• 

  Repeat Zero - the queue clears itself

• 

  Pain point

• 

  Time savings when using the app

• 

  snapshot 1

• 

  snapshot 2

• 

  snapshot 3

• 

  Architecture

Inspiration

Most support tickets repeat Support leads drown in password resets while novel incidents wait.

What it does

RepeatZero ingests tickets, retrieves runbooks, classifies with Nemotron Super, grounds repeats via Tavily, drafts replies, and applies a deterministic send-or-escalate gate with a cost ledger.

How we built it

TypeScript pipeline and React console in one Docker image on Render. Token Factory serves Nemotron Super and Nano; Tavily grounds citations; SSE streams typed envelopes; Playwright proves it.

Challenges we ran into

Two bugs hid behind green tests: ledgers emptied by a trace-id mismatch, and a corpus never loaded in-server. A live harvest on the real deploy exposed both; each fix verified end to end.

Accomplishments that we're proud of

Measured live: a repeat auto-sent at 0.90 confidence with five citations, a novel escalated with context, about a cent a ticket — plus a finished 2:59.9 narrated video.

What we learned

Metering must share the run's trace id or ledgers read empty. An unloaded corpus retrieves nothing. Golden caches mask missing keys. The gate earns trust: degraded input escalates, never sends.

What's next for RepeatZero — clearing the repeating bulk from queue

Keep the demo live, improve it after judging phase, ship the nightly re-index job, exercise Nano pings in real traffic, then fold every measured lesson into the next build.

Built With

• docker
• nebius
• nemotron
• node.js
• nvidia
• playwright
• pnpm
• react
• render
• sse
• tavily
• token-factory
• typescript
• vite
• vitest